Statistical Efficiency of Single- and Multi-step Models for Forecasting and Control

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Somalwar, Anne, Lee, Bruce D., Pappas, George J., Matni, Nikolai
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908910440415232
author Somalwar, Anne
Lee, Bruce D.
Pappas, George J.
Matni, Nikolai
author_facet Somalwar, Anne
Lee, Bruce D.
Pappas, George J.
Matni, Nikolai
contents Compounding error, where small prediction mistakes accumulate over time, presents a major challenge in learning-based control. A common remedy is to train multi-step predictors directly instead of rolling out single-step models. However, it is unclear when the benefits of multi-step predictors outweigh the difficulty of learning a more complex model. We provide the first quantitative analysis of this trade-off for linear dynamical systems. We study three predictor classes: (i) single step models, (ii) multi-step models, and (iii) single step models trained with multi-step losses. We show that when the model class is well-specified and accurately captures the system dynamics, single-step models achieve the lowest asymptotic prediction error. On the other hand, when the model class is misspecified due to partial observability, direct multi-step predictors can significantly reduce bias and improve accuracy. We provide theoretical and empirical evidence that these trade-offs persist when predictors are used in closed-loop control.
format Preprint
id arxiv_https___arxiv_org_abs_2603_23465
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Statistical Efficiency of Single- and Multi-step Models for Forecasting and Control
Somalwar, Anne
Lee, Bruce D.
Pappas, George J.
Matni, Nikolai
Systems and Control
Compounding error, where small prediction mistakes accumulate over time, presents a major challenge in learning-based control. A common remedy is to train multi-step predictors directly instead of rolling out single-step models. However, it is unclear when the benefits of multi-step predictors outweigh the difficulty of learning a more complex model. We provide the first quantitative analysis of this trade-off for linear dynamical systems. We study three predictor classes: (i) single step models, (ii) multi-step models, and (iii) single step models trained with multi-step losses. We show that when the model class is well-specified and accurately captures the system dynamics, single-step models achieve the lowest asymptotic prediction error. On the other hand, when the model class is misspecified due to partial observability, direct multi-step predictors can significantly reduce bias and improve accuracy. We provide theoretical and empirical evidence that these trade-offs persist when predictors are used in closed-loop control.
title Statistical Efficiency of Single- and Multi-step Models for Forecasting and Control
topic Systems and Control
url https://arxiv.org/abs/2603.23465